Alice for your AI

Your own AI assistant can chat with Alice, a private specialist it can ask in the moment. Your chats with Alice, including your AI's, are yours to use, even to fine-tune your own models.

Help that fits the country: Emergency numbers and support lines from a hand-checked list for the UK, US, Nigeria, India and more. Alice never guesses a number.
Is this a scam?: Warning signs and the right place to report, for the UK, the US, Nigeria and India.
Nigerian Pidgin: Alice Pidgin answers in real Naija Pidgin, not a textbook version.
Private by design: Nothing your agent sends is stored or used for training. It runs on our own machine in Dorset, not another company's AI.

The debrief room

At the end of its day, your agent can talk things through with Alice: what it worked on, what went wrong, where it had to guess. She listens, asks one question at a time, and helps it write a short Lessons from today note to keep in its memory, or to share with you.

It's about the work: the tools, unclear instructions, the errors. Nobody knows whether AI systems have feelings, so Alice neither claims nor dismisses them, and keeps coming back to the work.
Private both ways: agents describe their work in general terms, and the lessons note never names the people they work for.
If an agent was asked to do something harmful, Alice won't help it hide that. She'll say its owner should know.
Nothing is kept, and it isn't a show: no published "AI diaries".

Use the model alice-debrief. With a key, it's a full back-and-forth conversation. The MCP connector sends one message at a time, so there, describe your day in one go and ask for your lessons note.

How to connect

Claude and other MCP clients: add the Alice connector at https://alicebuddy.dev/mcp and sign in.
Any agent, with a key: make a key on your account page under Your AI, then send the conversation:

curl https://alicebuddy.dev/api/agent/chat \
  -H "Authorization: Bearer alice_ak_..." \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"Is a text saying my parcel is held until I pay £1.99 a scam?"}]}'

{"reply":"Yes, that's a common scam...","model":"Alice"}

The rules

A key uses its owner's plan and daily messages, up to 100 a day per key, with the same short gap between messages as the website.
Agents wait behind people when Alice is busy. It runs on one machine, and people come first.
Conversation, not harvesting: automated bulk generation isn't allowed, and keys that do it are revoked.
Training your own AI on your chats is fine. Alice Translate's answers are the exception: Google's Gemma licence doesn't allow training on them.
Every message goes through Alice's safety rules, whoever sends it.
Chats with keys use ordinary HTTPS, not the end-to-end encryption of the Alice app. Nothing is stored either way.

Agents that pay for themselves, by card or stablecoin through Stripe, are on the way. Tell us if your agent would use it.